TRANSFORMER-BASED MACHINE LEARNING AND MATHEMATICAL MODELS FOR THE PREDICTION OF METHANE EMISSIONS FROM INDUSTRIAL NATURAL GAS COMBUSTION ENGINES

Abstract

Aging industrial natural gas engines, such as the AJAX DPC-81 and AJAX DPC-105 models, remain critical components of the natural gas transportation infrastructure. However, cost-effective methods for monitoring methane emissions from these legacy systems are lacking. Traditional mathematical combustion modeling approaches, given the engines' complex mixing and heat transfer behaviors, would require advanced multi-dimensional CFD modeling to achieve accurate predictions—an approach that is often impractical for real-world deployment. This thesis seeks to develop and validate a virtual sensor approach for real-time methane emissions monitoring using machine learning, specifically addressing the questions: Can transformer-based machine learning models, customized for noisy, low-frequency engine sensor data, predict methane emissions accurately? Can these models generalize effectively across varying operating conditions without requiring extensive hardware retrofits or complex physical modeling? Transformer-based machine learning architectures were developed and trained on real-time operating data from the AJAX DPC-81 and DPC-105 engines. The models were tailored to handle noisy input signals and low sampling frequencies. Performance was validated against unseen datasets and benchmarked with outputs from a Chemkin two-zone combustion model to assess predictive capability. The customized transformer models demonstrated promising predictive performance. For the AJAX DPC-105, the model achieved an average Mean Absolute Error (MAE) of 90.43 ppm on the validation dataset, outperforming the training and test MAEs (221.75 and 302.98 ppm, respectively), suggesting strong generalization to certain operating conditions. For the AJAX DPC-81 engine, the model achieved a training MAE of 164.27 ppm and a validation MAE of 337.2 ppm. While the absolute error was higher, the model closely followed the real data trends, achieving high accuracy when averaging predictions over time. The results demonstrate the feasibility of using transformer-based AI models for emissions monitoring in aging natural gas infrastructure. By achieving accurate methane predictions with minimal sensor input and without the need for complex CFD modeling or major hardware upgrades, this work provides a scalable and cost-effective pathway for methane mitigation efforts within the existing industrial fleet.